Descriptions:
Google has released Gemma 4, a family of four open-weight models that represents a meaningful inflection point in the open-source AI landscape. For the first time, Google is shipping these models under a genuine Apache 2.0 license — no custom restrictions, no clauses prohibiting competition — meaning developers can fine-tune, modify, and commercialize them without legal friction. Sam Witteveen, who has tracked the Gemma line since its original release, frames the license change as Google finally playing by the same rules as Llama and Qwen, and notes the timing: several Chinese open-model providers are simultaneously pulling back from open releases.
The model family splits into workstation and edge tiers. The workstation group includes a 26B mixture-of-experts model (4B active parameters, derived from 128 experts) and a 31B dense model with a 256K context window and a new vision encoder with native aspect-ratio processing. Both carry architecture improvements traced back to Gemini 3 research. The edge models — E2B and E4B — are designed for phones, Raspberry Pis, and Jetson Nanos, and notably include a built-in ASR audio encoder compressed from 681M to 305M parameters, with frame duration cut from 160ms to 40ms for lower latency.
All four models ship with native vision, function calling, and long chain-of-thought reasoning built into the architecture rather than added as afterthoughts, with claimed strong results on MMMU Pro and SWE-Bench Pro. Google is positioning the workstation models as local coding assistants and IDE copilots, while the edge models target on-device speech recognition and real-time translation use cases.
📺 Source: Sam Witteveen · Published April 02, 2026
🏷️ Format: News Analysis
